Matlab implementation of our "Bayesian Zero-Shot Learning" paper. Accepted to ECCV 2020, TASK-CV Workshop. Authors: Sarkhan Badirli, Zeynep Akata, and Murat Dundar
Paper at: https://arxiv.org/pdf/1907.09624.pdf
We propose a hierarchical Bayesian model based on the intuition that actual classes originate from their corresponding local priors, each defined by a meta-class of its own. We derive the posterior predictive distribution (PPD) for a two-layer Gaussian mixture model to effectively blend local and global priors with data likelihood. These PPDs are used to implement a maximum-likelihood classifier, which represents seen classes by their own PPDs and unseen classes by meta-class PPDs. Across seven datasets with varying granularity and sizes, in particular on the large scale ImageNet dataset, we show that the proposed model is highly competitive against existing inductive techniques in the GZSL setting.
The code was implemented in Matlab. Any version greater 2016 should be fine to run the code.
You can download the datasets used in the paper from Google Drive. Create a data folder in your main project path and put the data under this folder.
To reproduce the results from the paper, open the Demo.m script and specify the dataset and model version (unconstrained or constrained). Please change the datapath to your project path in Demo.m script.
If you want to perform hyperparameter tuning, please comment out relevant sections from Demo.m script.
The results may vary 1-2% or less between identical runs in constrained model due to random initialization.
Feel free to drop me an email if you have any questions: [email protected]